Accuracy of least squares designed spatial FIR filters for segmentation of images of fluorescence stained cell nuclei.

A method for accurate, real-time image segmentation is needed for the development of a fully automated image cytometer that combines the speed and case-of-use of flow cytometry with the detailed morphometry of imaging. Object intensity variation and inherent optical blur make real-time segmentation challenging. The best spatial finite impulse response (FIR) filter, implemented as a convolution, was tested for sharpening edges and creating the required contrast. The filter and threshold segmentation steps were treated as a two-category linear classifier. Best 3 x 3 through 25 x 25 filters were designed utilizing the perceptron criterion and nonlinear least squares, and tested on ten montage images of a combined 1,070 manually segmented DAPI stained cell nuclei. The resulting image contrast, or class separation, led to simple automatic thresholding via the histogram intermodal minimum. Image segmentation accuracy began to plateau at 7 x 7 filters and did not increase above 15 x 15. Little loss in accuracy occurred with application to the images not used for design. This segmentation method provides a systematic, fast and accurate means of creating binary object maps useful for subsequent measurement, processing and cell classification.